kaldi-gstreamer-server

Why we included this project

Most Kaldi deployments start and end with batch decoding, but this server tackles the harder job of keeping a recognizer live for real-time use. It takes audio in over websockets and streams back partial hypotheses while the speaker is still talking, the same interaction pattern behind voice typing and live captioning. The master/worker split keeps the coordination logic separate from the decode processes, so workers can run on other machines and be added as concurrent sessions grow. Silence-based segmentation handles long streams, and lattice rescoring with a larger language model tightens accuracy when you need it. Expect to compile Kaldi and its GStreamer plugin yourself and to live with Python 2-era dependencies, which makes this a good fit for teams that are already comfortable inside the Kaldi toolchain.